The objective of the presentation is to give an overview of the new open source Acumos AI project including:
Scope of the project and the Acumos AI community
User journey
Towards an AI unified platform
High level technical architecture
Design studio for service chaining
Automating Google Workspace (GWS) & more with Apps Script
Towards an AI unified platform using Acumos, OW2con'18, June 7-8, 2018, Paris
1. Insights on the use of
Artificial Intelligence methods
for network management and
control in the perspective of
softwarized 5G
May 2018
S. Gosselin, S. Ben Jemaa, I. Grida Ben Yahia
V. Lemaire, N. Perrot, S. Sénécal
2. 2 Orange Unrestricted
Agenda
Context and scope
Possible roles and functionalities of Artificial Intelligence for
control and management of future networks
Orange experience from an internal AI network use case analysis
4. 4 Orange Unrestricted
AI could feed various steps of MAPE-K operational loop, leading
ultimately to autonomic operation
MONITOR
Gathers data from one or
more sources in the
environment
ANALYSE
Normalises and uses
data to develop
understanding of the
situation
PLAN
Analyses one or more
possible courses of
action
EXECUTE
Carries out decisions and
feeds back new data
MAPE-K : Monitor/Analyse/Plan/Execute - Knowledge
5. 5 Orange Unrestricted
Possible roles and
functionalities of Artificial
Intelligence for control
and management of
future networks
6. 6 Orange Unrestricted
Possible roles of AI for control and management of networks
AI specific role for a given use case has to be tuned depending on many factors:
• use case maturity, MAPE-K loop complexity and time scale, data availability and
reliability, centralized vs distributed implementation, operational impact, ….
AI allows a better
understanding of the
situation by operational
teams
Decisions are taken by
operational teams
Knowledge creation
AI proposes one or several
recommended actions
following situation analysis
Decisions are taken by
operational teams
Decision support
AI makes decisions that are
automatically carried out
No human involvement in
the MAPE-K loop
Decision making
7. 7 Orange Unrestricted
Data is the fuel for relevant use of AI in networks
Irrelevant / unreliable data AI will not help at all!
Data sources1
Data preparation (cleaning / filtering / pre-processing)2
Analysis / modelling3
Knowledge creation / Decision support / Decision making4
Action / adaptation5
MAPE-K
loop
8. 8 Orange Unrestricted
A plenty of AI methods, but a more limited number of functionalities
Predicting the value of a target variable for the futureData forecasting
Deriving statistical characteristics of dataData description
Finding the function linking target numerical variables with input variablesData regression
Determining possible drifts in data characteristicsVariation detection
Finding the function linking target categorical variables with input variablesData classification
Grouping of data into homogeneous clustersData segmentation
Discovering interesting relations between variablesData association
Identifying items which do not conform to an expected patternAnomaly detection
Controlling an interactive system or environment
Sequential optimi-
zation of parameters
10. 10 Orange Unrestricted
Classification of possible AI use cases: maturity level
Research
The use case is
investigated by
theoretical and
simulation
studies
Datasets do not
necessarily come
from the field
Operational
The use case is
already
implemented in
the field
Still
improvements of
AI method and its
implementation
can be sought
Proof of Concept
(PoC)
The use case has
been explored
with real
operational data
on a limited
scope
OR was run on a
real software /
network
infrastructure
Trial
The use case has
been explored
with real
operational data
on a limited
scope
AND was run on
a real software /
network
infrastructure
11. 11 Orange Unrestricted
General comments and insights from AI network use case analysis
Many use cases target a specific application of AI and need
specific datasets collected from the network, or even based on
external data
Many of these use cases have already led to Proofs of Concept
But only few use cases are at trial and operational stages
A large panel of possible applications of AI to networks is already
covered
Requirement to develop and improve internal runtime data platforms and external
design platforms (e.g. Acumos) for more “industrial” experimentation of AI on actual
operational data
The analysis has been limited to some Orange use cases, but:
AI applications to networks could represent high stakes for all the actors, operators,
vendors, GAFAM and third parties in general
The business model and ecosystem for specific applications of AI (e.g. to networks)
has to be built
12. 12 Orange Unrestricted
Which types of AI use cases for network control and management
A large possible scope of AI use cases for network management in general
Resource, service and even customer management
Various self-x functions, in particular self-optimization, self-diagnosis / healing /
protection, self-configuration
Fast and "simple" AI models could also help improving node-level or
function-level control mechanisms
e.g. performance congestion control, predictive scheduling
Positioning of AI use cases in Open Network Automation Platform (ONAP)
Data Collection, Analytics and Events (DCAE) to embed knowledge creation and
decision support use cases
Closed Loop Automation Management Platform (CLAMP) to embed decision making
use cases
Possible ultimate role of AI also in the update of policies as well as service
orchestration
13. 13 Orange Unrestricted
Implementation feasibility of AI use cases for network control and
management
Implementation / actuation of AI outcomes or decisions is not related to AI
techniques and data processing by themselves
Implementation / actuation done by the automation framework (e.g. ONAP) put in place for
network operations
The time dimension could provide additional constraints for decision actuation in the network,
especially for fast loops, i.e. control mechanisms
Still major general hindrances to implementation of AI use cases in networks
Dependence on data, lack of labelled data, readability and debuggability, …
Difficulty to match business objectives, optimization objectives of the AI model, and relevant
data
Dedicated software platforms with suitable APIs are necessary
To design, train, deploy and compose the machine learning/AI-based models
Such AI software platform(s) (e.g. Acumos) have to be combined / plugged with the runtime
automation framework (e.g. ONAP) for effective use case implementation